Low-rank extended Kalman filtering for online learning of neural networks from streaming data

TL;DR

This paper introduces a low-rank extended Kalman filtering method for online neural network learning from streaming data, offering a deterministic, efficient, and adaptive Bayesian inference approach that improves learning speed and responsiveness.

Contribution

The paper presents a novel low-rank plus diagonal EKF-based algorithm for online neural network parameter estimation, eliminating the need for step-size tuning and enhancing adaptation to non-stationary data.

Findings

Faster and more sample-efficient learning compared to stochastic variational inference.

Improved adaptation to changing data distributions.

Enhanced reward accumulation in contextual bandit applications.

Abstract

We propose an efficient online approximate Bayesian inference algorithm for estimating the parameters of a nonlinear function from a potentially non-stationary data stream. The method is based on the extended Kalman filter (EKF), but uses a novel low-rank plus diagonal decomposition of the posterior precision matrix, which gives a cost per step which is linear in the number of model parameters. In contrast to methods based on stochastic variational inference, our method is fully deterministic, and does not require step-size tuning. We show experimentally that this results in much faster (more sample efficient) learning, which results in more rapid adaptation to changing distributions, and faster accumulation of reward when used as part of a contextual bandit algorithm.